Repository navigation
Bootstrap kept-tree rollback can leave no tree, and a hand-run report mislabels its commit #1770
Description
Two narrow gaps in the kept-tree handling the #1756 change added to host-setup/bootstrap.sh, bootstrap.ps1, and scripts/skills_install.py.
The rollback can leave no tree
The swap moves the old tree aside to .old and then renames .new into place. If that second rename fails, the swap renames .old back. If the rollback rename fails as well, or an interrupt lands between the two renames, the exit cleanup deletes both .new and .old. Nothing is then left at the tree's name, and the plugin loads nothing. In bash, a failing rollback mv trips set -e. In PowerShell, a throwing rollback Move-Item escapes the catch into finally.
- Fix: where the tree path is absent, cleanup renames an owned
.oldback rather than deleting it.
A hand-run report mislabels the commit's source
intended_commit labels an archive commit's ref SKILLS_SOURCE_COMMIT. Since #1756, source_ref also reads the commit from the tree's .bootstrap-commit marker. A --report run by hand from a kept tree, with that variable unset, therefore names a variable nobody set.
- Fix: label the ref by whichever source supplied the commit.
Found by the local strict review on the #1756 change, and deferred at the maintainer's stop point.
Activity
ptr727 commentedon Sep 24, 2026
A third gap of the same kind, found by the local strict review on #1773.
bootstrap.shremoves an owned tree withrm -rf, which can delete the.bootstrap-ownedmarker before the rest of the tree. If that removal stops part way, for example on a signal or on one entry that fails, the leftover directory carries no marker. The loader then no longer counts it as its own. Every later swap dies with "exists and this loader did not create it ... Choose another --dir", and that remedy is wrong: the fix is to delete the leftover.tree(the oldremove_tree), and Bootstrap registers its temporary tree as the live Claude Code skills source, then deletes it #1756 adds the same pattern for.old.bootstrap.ps1already avoids it: itsRemove-Retiredremoves the marker last.ptr727 commentedon Sep 24, 2026
Two more of the same kind, both predating #1756, found by the local strict review on #1773:
--report --intended REVrun from a bootstrap-extracted tree still asks git in that tree. When a revision is named,intended_commitskips the archive short-circuit. If the tree sits under a home directory kept in git, the enclosing repository resolvesREV, and the report calls the copy stale against a commit from an unrelated repository. The loaders never pass--intended, so only a run by hand reaches this.--intendedthere.--reportexits 1 with "cannot be judged", and the loader then prints "missing or stale, and --host or --skills lands it".scripts/README.mdsays a re-install does not fix that state.reasonrather than by the exit code alone.ptr727 commentedon Sep 24, 2026
Two more from the local strict review on #1773, both on
bootstrap.ps1'sInvoke-Cleanupin the double-failure case, where a swap's rename fails and so does its rollback.Remove-Item -Recurse. The lock that makes the swap's rename fail, such as an antivirus scan of the freshly extracted tree, also makes this removal throw part way. That skips the restore and its "move it by hand" warning. The removal can also delete.bootstrap-ownedfirst, leaving a staging tree the next run refuses as not its own.Remove-Retiredremoves the old tree. This is the same fix as the bash marker-order point above, applied to every removal of an owned tree.ptr727 commentedon Sep 24, 2026
One more, from Copilot on #1773:
remove_ownedandis_ourstest[[ -e ]], so a dangling symlink at a managed path such asskills-tree.newreads as absent.mkdir -pthen fails with a generic "File exists". Nothing is removed. Fix: treat-Las existing in the ownership check, and refuse with the loader's own message.ptr727 commentedon Sep 24, 2026
One more, from CodeRabbit on #1774 (
bootstrap.ps1Invoke-SwapIn): whenskills-treeis absent and only a lockedskills-tree.oldremains, the swap first tries to remove.old. That fails on the lock, so the new staging tree is never promoted, and the registered name stays empty. Fix: skip the pre-removal of.oldwhen the tree name is empty. Keep the refusal when both exist, since.oldhas to be free before the live tree moves aside.Promote the Bootstrap's Kept Skills Tree to Main (#1774)
Close the Bootstrap Kept-Tree Removal and Report Gaps (#1794)
Promote the Bootstrap Kept-Tree Fixes and the HomeAutomation-Config A…